Diffusion model iteration correction precipitation prediction method based on space-time conditions

Abstract

The invention discloses a diffusion model iteration correction precipitation prediction method based on space-time conditions, which improves the precision and timeliness of precipitation prediction, and is especially suitable for prediction of extreme weather events. The method comprises the following steps: firstly, receiving multi-modal input data including a radar echo image, a time identifier, earth surface remote sensing data and multichannel data of a meteorological satellite, and extracting characteristics of various input data by utilizing a plurality of encoders; in order to solve the problem that radar image rainfall distribution does not conform to a probability model, after an encoder is adopted to extract features, M frames of radar images are classified into different categories through a k-means clustering algorithm, so that corresponding prediction models are selected according to the categories, and multi-task prediction is carried out. A diffusion model predicts precipitation in a hidden space in a step-by-step generation mode, and iterative correction is introduced in each time period of a prediction result so as to reduce error accumulation. And finally, outputting a high-precision long-time sequence rainfall prediction result by the model. According to the method, multi-modal information, time conditions and an iterative correction structure are combined, the rainfall prediction capacity under complex terrains and extreme weather is remarkably improved, and the method has high practical value and application prospects.

Description

Technical Field

[0001] The invention discloses a precipitation prediction method based on iterative correction of diffusion model under spatiotemporal conditions, and in particular relates to a precipitation prediction method using a denoised diffusion probability model under spatiotemporal conditions to predict radar echo reflection images. Background Art

[0002] In the existing precipitation prediction methods, traditional models are difficult to accurately capture the spatiotemporal dynamic characteristics of precipitation, especially in extreme precipitation events, and the prediction results are often biased. In recent years, deep learning models have shown great advantages in spatiotemporal series data modeling, but due to the limited nonlinear representation ability of traditional deep learning models in complex dynamic systems, their accuracy and stability in precipitation prediction still need to be improved. Therefore, it is of great significance to propose an effective iterative correction method for precipitation prediction.

[0003] Denoising Diffusion Probabilistic Model (DDPM). The DDPM diffusion model is an excellent image generation model with good image quality and generation capabilities. Existing precipitation prediction based on DDPM usually uses radar echo reflection images or multi-channel data from meteorological satellites. This type of method uses a diffusion model to gradually generate or repair images to achieve the purpose of precipitation prediction. Radar images can provide local precipitation intensity and distribution information, while multi-channel data from meteorological satellites can capture a wider range of meteorological changes, such as temperature, humidity, cloud thickness, etc. The combination of this information can improve the accuracy of precipitation prediction.

[0004] Existing deep learning methods face multiple challenges in precipitation prediction. First, these models have limitations in capturing spatial details, especially in high-resolution radar images, and it is difficult to accurately identify the detailed features of local high-intensity precipitation, resulting in deviations in the prediction of precipitation intensity and distribution. Second, precipitation is a highly spatiotemporally correlated process, and existing models perform poorly in accurately modeling precipitation temporal dynamics (such as the start, intensity change, and end of precipitation), especially in extreme weather events such as heavy rain and severe convective weather. In addition, the precipitation process also involves a large number of nonlinear features, including the multiple effects of terrain, temperature, and wind speed. Although methods such as LSTM and Transformer have certain advantages in capturing temporal nonlinear relationships, it is still difficult to effectively model these complex nonlinear relationships in precipitation prediction across time scales. Summary of the invention

[0005] In order to overcome the defects of existing models in precipitation prediction methods, the present invention proposes a variety of prediction techniques based on DDPM. First, temporal and spatial conditions are added to the input. Second, a decomposition-multitask method is used to accurately predict radar images of different precipitation intensities. Third, an iterative corrector method is used simultaneously to increase the performance degradation problem in long-term prediction.

[0006] Existing models also use multimodal radar data and satellite multi-channel data, but when dealing with complex terrain, the accuracy of heavy precipitation forecast and warning is still low, and the lead time is insufficient. Therefore, the present invention introduces satellite remote sensing images to obtain surface morphological features, thereby improving the ability to predict precipitation in complex terrain areas.

[0007] Precipitation is also significantly affected by time factors. For example, China's 24 solar terms summarize the pattern of rainfall in a specific time period. Therefore, we divide the 24 hours of a day into four time periods and multiply the number of days in a year by four as a time condition (called a time mark) to assist prediction.

[0008] The formation of precipitation is a complex natural process, involving multiple factors such as the condensation of water vapor in the atmosphere, the formation of clouds, and the movement of air currents, which leads to the complexity of radar images. Based on this, the present invention proposes a "decomposition-multitask" prediction technology, which regards the precipitation image as multiple components and decomposes it into areas with different rainfall intensities. An independent prediction task is designed for each area to more accurately capture and reflect the precipitation characteristics.

[0009] In order to solve the problem of insufficient time advance of traditional models, this paper proposes a scheme combining iterative prediction with a corrector. Specifically, a prediction model based on DDPM is adopted, and the prediction result of each step is used as the input of the next model, so as to gradually enhance the time span of the prediction. For the accumulation of prediction errors, we further introduce DDPM for correction to output the final accurate prediction results.

[0010] Current deep learning methods are deficient in building dependencies between conditional information and prediction results. To address this problem, the present invention has specially designed the model structure, introduced a causal attention mechanism into the Transformer backbone network, and established spatiotemporal queries between conditional information and prediction results, so that the model can effectively capture long-term dependencies. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is the overall framework diagram of the precipitation prediction based on the diffusion model of multimodal spatiotemporal conditions.

[0012] Figure 2 This is a schematic diagram of the Transformer network structure.

[0013] Figure 3 It is the training flowchart of the iterative correction method.

[0014] Figure 4 It is the reasoning process of the iterative correction method.

[0015] Figure 5 This is a diagram of the multi-model selection architecture based on radar precipitation intensity distribution. DETAILED DESCRIPTION

[0016] The present invention provides a precipitation prediction method based on diffusion model iterative correction of spatiotemporal conditions. This method uses spatiotemporal condition denoising diffusion probability model (DDPM) to accurately predict precipitation, which can better capture complex spatiotemporal dynamic characteristics, especially in extreme weather conditions. Figures 1 to 5 Embodiments of the present invention will be described in detail.

[0017] Figure 1 The overall structure of the present invention is shown. The system first receives multiple observation images (such as radar echo images), passes through multiple encoders (such as E 1 , D 1 ,…,D a ) to extract preliminary features and transform them into latent space. The diffusion process starts from the initial state Z 0 At the beginning, after multiple layers of iteration, it is updated to Z T , to generate the final prediction results. In the hidden space, we introduced the Transformer network, which uses multi-step iterations to gradually generate and optimize the prediction results. The right side of the figure shows the conditional network, which inputs multimodal information (such as time stamps, surface remote sensing images, meteorological satellite multi-channel data, etc.) into multiple encoders (E 2 、E 3 、E 4 、E 5 ) is processed and then feature fused to generate high-dimensional conditional features for precipitation prediction. The introduction of these conditional information can enhance the adaptability of the model to complex terrain and time changes, thereby improving the accuracy of precipitation prediction.

[0018] Figure 2 The Transformer network structure is described in detail. Each layer contains normalization (Norm), multi-head self-attention (MSA) and multi-layer perceptron (MLP) modules, which extract deep features by processing the input data layer by layer. In each Transformer module, the input first passes through the causal self-attention layer. The causal self-attention layer is a kind of MSA, which performs standard self-attention calculations on the features of the observed image as V and K and the conditional fusion feature as Q, that is: Q=XW Q , K = XWK , V = XW V (1) Then it is processed continuously through three TB blocks. Each TB block includes the use of a normalization layer to improve numerical stability, and then a multi-head self-attention mechanism to capture the long-distance dependencies between input data, and finally a multi-layer perceptron for further nonlinear transformation. By stacking multiple Transformer modules, the model can effectively extract the multi-scale features required for precipitation prediction, thereby enhancing the ability to model complex spatiotemporal relationships.

[0019] Figure 3 The training process of the iterative correction method is shown, where Figure 3 It’s the training process. Figure 4 It is a process of reasoning. Figure 3 The horizontal direction represents the time axis (or radar frame). 0 is the reference point of the current time, T -3:0 Indicates that from the current time T 0 Push back 3 frames, including 4 frames of images in total; T 1:20 Indicates that from time point T 1 to T 20 , a total of 20 frames, and so on. Figure 3 During the training process, 24 consecutive frames are combined into sample pairs, and the first 4 frames (such as T -3:0 ) plus satellite multi-channel data, time stamp, and surface remote sensing images as conditions, and the last 20 frames (such as T 1:20 ) as the observation data of the diffusion model to train the model. Then iterative correction training is performed, that is, the last 4 frames (such as T 17:20 ) and other conditions as the conditional input of the second model, and T 21:40 The observation data is used to train the correction model 1. The subsequent correction model training is similar. Figure 4 A total of 1 prediction model and 3 correction models need to be trained. Both the prediction model and the correction model are based on Figure 1 The network structure shown.

[0020] Figure 4 This is the reasoning process of the iterative correction method. In the reasoning process, also from the time point T -3:0 Starting from the 4 frames of data, we first generate the preliminary prediction result T 1:20 In each prediction time period, take the last 4 frames predicted in the previous stage (for example, T 17:20 ) predict 20 frames (for example, T 21:40), and the corresponding correction model (consistent with the model in the training process) is applied to adjust the current prediction to further optimize the prediction result. Through this iterative correction method, the present invention can significantly improve the accuracy of long-term precipitation prediction, especially in the scenario of extreme precipitation events, with higher stability and robustness. Figure 4 In total, the joint prediction of 1 prediction model and 3 correction models is shown, which can predict 80 frames of images. If the interval between each frame is 5 minutes, the next 400 minutes can be predicted; if the interval between each frame is 10 minutes, the next 800 minutes can be predicted.

[0021] Figure 5 The multi-model selection architecture based on feature extraction and clustering is presented. First, the encoder E 1 Features are extracted from M frames of radar images, which contain the distribution information of precipitation. Then, the extracted features are connected in series and the images are classified into five categories using the k-means clustering algorithm. According to the clustering results, the corresponding prediction model is selected (see Figure 1 The multi-task prediction model in the prediction model is used to adapt to different precipitation intensity distribution scenarios. According to the precipitation distribution characteristics of different categories, the corresponding models are trained and optimized respectively to improve the generalization ability and prediction accuracy of the model.

[0022] The detailed implementation steps are as follows:

[0023] Step 1: Multimodal input feature extraction and fusion The system first receives input from multiple data sources, including N frames of radar echo images, time stamps, surface remote sensing data, and multi-channel data from meteorological satellites. These inputs are processed by multiple encoders:

[0024] Radar Image Encoder E 1 Extract the basic features of M-frame radar images and provide initial precipitation distribution information.

[0025] Conditional Encoder E 2 To E 5 Different data types are processed separately, including time stamps, surface features, and satellite multi-channel data, and various types of data are converted into a unified feature representation. The extracted multimodal features are fused to generate high-dimensional features for subsequent precipitation prediction. This fusion process enables the model to combine multiple environmental information and improve the prediction accuracy under complex terrain and climate conditions.

[0026] Step 2: Feature extraction and clustering classification of M-frame radar images The precipitation distribution information of M-frame radar images does not belong to any known probability model, so it is extracted through encoder E 1After extracting features, these features are concatenated and classified using the k-means clustering algorithm. The clustering results divide the M frames into five categories, so that scenes with different precipitation intensity distributions can be identified. This feature-based classification makes the subsequent model selection more targeted.

[0027] Step 3: Multi-task model selection and training According to the clustering results, select the corresponding prediction model (such as Figure 1 Specifically, each category corresponds to a specially trained prediction model to adapt to the precipitation distribution characteristics of that category. First, the four frames of images are classified, and then the corresponding models are selected and trained according to different categories. This classification-based multi-model architecture improves the generalization ability of the model and can more accurately respond to different types of precipitation situations.

[0028] Step 4: The iterative prediction process of the diffusion model starts from the initial state Z in the latent space 0 At the beginning, the diffusion model gradually generates precipitation forecast results. In this process, the features in the latent space are processed by a multi-layer Transformer network, and the deep features are extracted layer by layer through the multi-head self-attention mechanism and multi-layer perceptron to capture the long-term dependencies between spatiotemporal features. With the iteration of the diffusion process, the prediction accuracy of the model continues to improve, and finally outputs high-quality precipitation forecast results.

[0029] Step 5: Iteratively correct the model for long time series prediction optimization. The prediction model is first optimized at the initial time point (e.g., T -3:0 ) Generate the first prediction result T 1:20 , then in each time period (such as T 21:40 ,T 41:60 ,T 61:80 ) Iterative correction prediction is gradually performed. For the prediction results of each time period, the correction model adjusts the error to reduce the accumulation of prediction errors. The corrected prediction results are recursively input into the prediction model of the next time period to form an iterative correction structure, which effectively extends the prediction time span and improves the prediction stability, especially in extreme precipitation events, which can significantly reduce errors.

[0030] Step 6: Multi-model output of multi-task parallel processing architecture is as follows Figure 5 As shown in the figure, the system selects a suitable prediction model based on the precipitation intensity distribution of M-frame radar images. After the input data is processed by the central module, it branches into three parallel modules, each of which corresponds to a specific category of prediction tasks. Each branch module predicts different types of precipitation intensity distribution, thereby improving the overall output level and prediction accuracy.

Claims

1. A precipitation prediction method based on iterative correction of diffusion model in spatiotemporal conditions, characterized in that: The following steps are involved: 1) Receive multi-modal input data, including N frames of radar echo images, time stamps, surface remote sensing data, and meteorological satellite multi-channel data; 2) using a plurality of encoders to extract features of the multimodal input data, wherein encoder E1 is used to extract features of M-frame radar images, and encoders E2 to E5 process time stamps, surface remote sensing data, and satellite multi-channel data, respectively; 3) After the extracted M-frame radar image features are connected in series, these images are classified into multiple categories using the k-means clustering algorithm; 4) Based on the clustering results, select the corresponding category prediction model to perform multi-task prediction on precipitation distribution characteristics of different categories; 5) In the latent space, the denoising diffusion probability model DDPM is used to gradually generate precipitation prediction results from the initial state; 6) Iteratively correct the prediction results in each time period and gradually input them into the next time period to form a recursive structure for long time series prediction; 7) Output the final accurate precipitation prediction results.

2. The precipitation prediction method based on the diffusion model iterative correction of spatiotemporal conditions according to claim 1 is characterized in that: In step 3), the k-means clustering algorithm is used to classify the features of the M-frame radar images into five categories.

3. The precipitation prediction method based on the diffusion model iterative correction of spatiotemporal conditions according to claim 1 is characterized in that: In the step 4), different categories of prediction models are trained for specific precipitation intensity distribution scenarios to improve the generalization and prediction accuracy of the models.

4. The precipitation prediction method based on the diffusion model iterative correction of spatiotemporal conditions according to claim 1 is characterized in that: The denoising diffusion probability model DDPM generates a step-wise diffuse prediction result in the latent space and performs iterative correction in multiple time periods to reduce error accumulation.

5. The precipitation prediction method based on the diffusion model iterative correction of spatiotemporal conditions according to claim 1 is characterized in that: The iterative correction in step 6) includes one prediction model and several correction models. The correction model is used to adjust the error of the last predicted frame to optimize the prediction result of a long time span.

6. The precipitation prediction method based on the diffusion model iterative correction of spatiotemporal conditions according to claim 1 is characterized in that: A causal self-attention mechanism is introduced into the backbone network Transformer of the diffusion model to establish spatiotemporal queries between conditional information and prediction results, enabling the model to effectively capture long-term dependencies.

7. The precipitation prediction method based on diffusion model iterative correction of spatiotemporal conditions according to claim 1 is characterized in that: The time mark includes time condition information based on the 24-hour daily cycle division and the annual solar terms to capture the temporal periodic changes of precipitation.

8. The precipitation prediction method based on diffusion model iterative correction of spatiotemporal conditions according to claim 1 is characterized in that: The surface remote sensing data contains terrain feature information to enhance the prediction accuracy of the model under complex terrain conditions.

9. The precipitation prediction method based on the diffusion model iterative correction of spatiotemporal conditions according to claim 1 is characterized in that: The meteorological satellite multi-channel data includes meteorological variables such as temperature, humidity and cloud thickness.

10. A precipitation prediction system based on spatiotemporal conditions, comprising: (1) Multimodal data receiving module, used to receive radar echo images, time stamps, surface remote sensing data and meteorological satellite multi-channel data; (2) a feature extraction module, used to extract features of multimodal input data; (3) A clustering classification module, used to perform clustering classification on the features of M radar images; (4) a multi-task model selection module, used to select the corresponding prediction model according to the clustering results; (5) De-noised diffusion probability model, used to generate precipitation forecast results; (6) an iterative correction module, used for iteratively correcting the prediction results; (7) Output module, used to output accurate precipitation prediction results.

Images